Artificial Intelligence-Based Cardiac Ultrasound Examination Assistance System

Abstract

This invention relates to the field of image processing technology, and more particularly to an artificial intelligence-based cardiac ultrasound examination auxiliary system, comprising: an image processing module, including an acquisition unit for acquiring dynamic cardiac ultrasound images using an ultrasound probe; a model training module connected to the image processing module, including a model generation unit for training an initial model based on optimized images to output a convolutional neural network model; a feature extraction module connected to the model training module, including a feature extraction unit for extracting key features of the cardiac structure through the convolutional neural network model to output features of the dynamic cardiac ultrasound image; and a control module connected to the image processing module, the model training module, and the feature extraction module respectively, for determining the pulse repetition frequency of the ultrasound probe based on the artifact occurrence rate of the dynamic cardiac ultrasound image. This invention improves the processing stability of dynamic cardiac ultrasound images.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an artificial intelligence-based cardiac ultrasound examination assistance system. Background Technology

[0002] In the current technology, with the rapid development of medical imaging technology, cardiac ultrasound (echocardiography) has become a core tool for the diagnosis, treatment assessment and follow-up of cardiovascular diseases. It achieves automatic segmentation of cardiac structures and disease classification through algorithms such as convolutional neural networks and Transformers, combines temporal modeling to analyze dynamic blood flow and valve motion, and enhances data diversity by using multimodal data fusion and adversarial generative networks. It also achieves real-time analysis by deploying lightweight models to portable devices, which solves the problems of strong operation dependence, high measurement subjectivity and artifact interference in traditional ultrasound examinations, and ultimately improves diagnostic efficiency and accuracy.

[0003] Chinese Patent Publication No. CN112233087A discloses an artificial intelligence-based ophthalmic ultrasound disease diagnosis method and system, including: a client and a server; the client is used to monitor and acquire images from ophthalmic ultrasound examination equipment, and upload them to the server via the network, receiving and displaying the current eyeball position, lesion position, and disease type in real time; the server is used to receive ophthalmic ultrasound images acquired from the client, perform eyeball segmentation using a convolutional neural network, identify whether there is a disease in the segmented eyeball, and if there is a disease, further identify the disease type and the location of the lesion, and feed back the various identification results to the client.

[0004] It is evident that existing technologies have the following problems: due to the presence of cardiac motion artifacts and blood flow signal interference in dynamic cardiac ultrasound images, existing systems lack effective mechanisms for handling these interfering factors. In addition, the training data does not cover enough cardiac disease scenarios, resulting in weak model generalization ability and thus the processing stability of dynamic cardiac ultrasound images does not meet the requirements. Summary of the Invention

[0005] To address this, the present invention provides an artificial intelligence-based cardiac ultrasound examination assistance system to overcome the problems in the prior art, such as cardiac motion artifacts and blood flow signal interference in dynamic cardiac ultrasound images. Existing systems lack effective mechanisms for handling these interference factors, and the training data does not cover enough cardiac disease scenarios, resulting in weak model generalization ability and thus the instability of processing dynamic cardiac ultrasound images does not meet the requirements.

[0006] To achieve the above objectives, the present invention provides an artificial intelligence-based cardiac ultrasound examination assistance system, comprising:

[0007] The image processing module includes an acquisition unit for acquiring dynamic cardiac ultrasound images using an ultrasound probe and a preprocessing unit connected to the acquisition unit for preprocessing the dynamic cardiac ultrasound images to output optimized images.

[0008] The model training module, which is connected to the image processing module, includes a model generation unit for training an initial model based on the optimized image to output a convolutional neural network model, and a model updating unit connected to the model generation unit for updating the convolutional neural network model using a stochastic gradient descent optimization algorithm.

[0009] The feature extraction module, which is connected to the model training module, includes a feature extraction unit for extracting key features of the heart structure through a convolutional neural network model to output dynamic features of cardiac ultrasound images, and an analysis unit connected to the feature extraction unit for analyzing the dynamic features of cardiac ultrasound images to output analysis results.

[0010] The control module, which is connected to the image processing module, the model training module, and the feature extraction module respectively, is used to determine the pulse repetition frequency of the ultrasound probe based on the artifact occurrence rate of the dynamic cardiac ultrasound image, or to determine the gradient clipping threshold based on the analysis accuracy of the convolutional neural network model, and to determine the transmission power of the ultrasound probe based on the signal-to-noise ratio of the dynamic cardiac ultrasound image.

[0011] Furthermore, the control module is used to determine whether the processing stability of the dynamic cardiac ultrasound image meets the requirements based on whether the artifact occurrence rate of the dynamic cardiac ultrasound image is less than or equal to a preset first occurrence rate.

[0012] The control module is used to determine that the processing stability of the dynamic cardiac ultrasound image does not meet the requirements if the artifact occurrence rate of the dynamic cardiac ultrasound image is greater than a preset first occurrence rate.

[0013] Furthermore, the control module is used to preliminarily determine that the training effectiveness of the convolutional neural network model does not meet the requirements based on the fact that the artifact occurrence rate of the cardiac ultrasound dynamic image is greater than the preset first occurrence rate and less than or equal to the preset second occurrence rate.

[0014] Furthermore, the control module is used to increase the pulse repetition frequency of the ultrasound probe based on the fact that the artifact occurrence rate of the cardiac ultrasound dynamic image is greater than the preset second occurrence rate.

[0015] The increase in the pulse repetition frequency of the ultrasound probe is determined by the difference between the artifact occurrence rate of the dynamic cardiac ultrasound image and a preset second occurrence rate.

[0016] Furthermore, the control module is used to determine that the training effectiveness of the convolutional neural network model meets the requirements based on the fact that the analysis accuracy of the convolutional neural network model is greater than or equal to a preset second accuracy.

[0017] The control module is used to determine that the training effectiveness of the convolutional neural network model does not meet the requirements if the analysis accuracy of the convolutional neural network model is less than a preset second accuracy.

[0018] Furthermore, the control module is used to reduce the gradient clipping threshold based on the fact that the analysis accuracy of the convolutional neural network model is greater than a preset first accuracy and less than a preset second accuracy.

[0019] Furthermore, the control module is used to preliminarily determine that the accuracy of extracting features from the dynamic cardiac ultrasound image does not meet the requirements if the analysis accuracy of the convolutional neural network model is less than or equal to the preset first accuracy.

[0020] Furthermore, the reduction in the gradient clipping threshold is determined by the difference between the analysis accuracy of the convolutional neural network model and a preset first accuracy.

[0021] Furthermore, the control module is used to determine that the accuracy of extracting features from the dynamic cardiac ultrasound image does not meet the requirements if the signal-to-noise ratio of the dynamic cardiac ultrasound image is less than a preset signal-to-noise ratio, and to increase the transmission power of the ultrasound probe.

[0022] Furthermore, the increase in the transmission power of the ultrasound probe is determined by the difference between the preset signal-to-noise ratio and the signal-to-noise ratio of the dynamic cardiac ultrasound image.

[0023] Compared with existing technologies, the beneficial effects of this invention are as follows: The system of this invention, by setting up an image acquisition module, a model training module, a feature extraction module, and a control module, adjusts the pulse repetition rate according to the artifact occurrence rate of dynamic cardiac ultrasound images. Utilizing artificial intelligence image recognition training, it quickly identifies morphological changes in the heart under a four-chamber ultrasound view during cardiac arrest, as well as common cardiac lesions. Simultaneously, it intelligently identifies images with non-standard cross-sections and prompts the scanner to adjust the scanning angle. For artifacts caused by air, bone, and other factors that affect the clarity and edge recognition of cardiac structures, increasing the duty cycle of the pulse repetition frequency can increase the effective signal emission per unit time, thereby enhancing the intensity and continuity of the ultrasound echo signal. This reduces image fluctuations caused by signal attenuation, especially in areas with greater detection depth or weak tissue echoes. Stable signal output can improve image contrast and resolution, reducing artifacts caused by signal instability. The gradient clipping threshold is adjusted based on the analysis accuracy of the convolutional neural network model. Due to insufficient generalization ability of the model, the convolutional neural network model may perform well on the training set. However, when faced with unseen real data, the accuracy and stability of feature extraction significantly decrease. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general patterns, leading to deviations in the extraction of the contours of fine cardiac structures. This, in turn, reduces the accuracy of the diagnostic assistance system. By reducing the gradient clipping threshold, excessively large gradients can be more strictly limited, suppressing the model's overfitting to noise and local features in the training data and preventing the model from misjudging occasional noise signals as key patterns. This reduces the deviation in the extraction of the contours of fine cardiac structures. Adjusting the transmission power of the ultrasound probe based on the signal-to-noise ratio of the dynamic cardiac ultrasound image is necessary. However, due to hardware limitations of the ultrasound equipment itself, image resolution is reduced, making fine cardiac structures such as chordae tendineae and valve edges blurry. Continuous and accurate spatiotemporal features cannot be obtained, making it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately resulting in a significant decrease in the accuracy of the diagnostic assistance system. Increasing the transmission power of the ultrasound probe can increase the initial energy of the ultrasound waves, enhancing the echo signal intensity of deep tissues and far-field structures, and improving image contrast.

[0024] Furthermore, the system of the present invention adjusts the pulse repetition rate by setting a preset first occurrence rate and a preset second occurrence rate. Utilizing artificial intelligence image recognition training, it quickly identifies morphological changes in the heart under a four-chamber ultrasound view during cardiac arrest, as well as common cardiac lesions. Simultaneously, it intelligently identifies images with non-standard sections and prompts the scanner to adjust the scanning angle. For artifacts caused by air, bone, and other factors that affect the clarity and edge recognition of the heart structure, increasing the duty cycle of the pulse repetition frequency increases the effective signal emission per unit time, thereby enhancing the intensity and continuity of the ultrasound echo signal. This reduces image fluctuations caused by signal attenuation, especially in areas with greater detection depth or weaker tissue echoes. Stable signal output improves image contrast and resolution, reduces artifacts caused by signal instability, and further enhances the processing stability of cardiac ultrasound images.

[0025] Furthermore, the system of the present invention adjusts the gradient clipping threshold by setting a preset first accuracy rate and a preset second accuracy rate. Due to insufficient generalization ability of the model, the convolutional neural network model may perform well on the training set, but when faced with unseen real data, the accuracy and stability of feature extraction will significantly decrease. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general rules, resulting in deviations in the extraction of the contours of the heart's fine structures. This leads to a decrease in the accuracy of the diagnostic assistance system. By reducing the gradient clipping threshold, excessively large gradients can be more strictly limited, suppressing the model's overfitting of noise and local features in the training data, and preventing the model from misjudging occasional noise signals as key patterns. This reduces the deviation in the extraction of the contours of the heart's fine structures and further improves the processing stability of cardiac ultrasound images.

[0026] Furthermore, the system described in this invention adjusts the transmission power of the ultrasound probe by setting a preset signal-to-noise ratio. Due to hardware limitations of the ultrasound equipment itself, the image resolution is reduced, and fine cardiac structures such as chordae tendineae and valve edges become blurred, making it impossible to obtain continuous and accurate spatiotemporal features. This makes it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately leading to a significant decrease in the accuracy of the diagnostic assistance system. By increasing the transmission power of the ultrasound probe, the initial energy of the ultrasound waves can be increased, thereby enhancing the echo signal intensity of deep tissues and far-field structures, improving image contrast, and further improving the processing stability of cardiac ultrasound images. Attached Figure Description

[0027] Figure 1 This is a block diagram of the overall structure of the artificial intelligence-based cardiac ultrasound examination assistance system according to an embodiment of the present invention;

[0028] Figure 2This is a logic flowchart illustrating the process of determining the pulse repetition frequency of an ultrasound probe in an artificial intelligence-based cardiac ultrasound examination assistance system according to an embodiment of the present invention.

[0029] Figure 3 This is a flowchart illustrating the process of determining the gradient clipping threshold in the artificial intelligence-based cardiac ultrasound examination assistance system according to an embodiment of the present invention.

[0030] Figure 4 This is a logic flowchart illustrating the process of determining the transmission power of an ultrasound probe in an artificial intelligence-based cardiac ultrasound examination auxiliary system according to an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0032] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0033] Please see Figure 1 , Figure 2 , Figure 3 as well as Figure 4 The diagrams shown are, respectively, an overall structural block diagram of the artificial intelligence-based cardiac ultrasound examination auxiliary system according to an embodiment of the present invention, a logic flowchart for determining the pulse repetition frequency of the ultrasound probe, a logic flowchart for determining the gradient clipping threshold, and a logic flowchart for determining the transmission power of the ultrasound probe. The present invention provides an artificial intelligence-based cardiac ultrasound examination auxiliary system, comprising:

[0034] The image processing module includes an acquisition unit for acquiring dynamic cardiac ultrasound images using an ultrasound probe and a preprocessing unit connected to the acquisition unit for preprocessing the dynamic cardiac ultrasound images to output optimized images.

[0035] The model training module, which is connected to the image processing module, includes a model generation unit for training an initial model based on the optimized image to output a convolutional neural network model, and a model updating unit connected to the model generation unit for updating the convolutional neural network model using a stochastic gradient descent optimization algorithm.

[0036] The feature extraction module, which is connected to the model training module, includes a feature extraction unit for extracting key features of the heart structure through a convolutional neural network model to output dynamic features of cardiac ultrasound images, and an analysis unit connected to the feature extraction unit for analyzing the dynamic features of cardiac ultrasound images to output analysis results.

[0037] The control module, which is connected to the image processing module, the model training module, and the feature extraction module respectively, is used to determine the pulse repetition frequency of the ultrasound probe based on the artifact occurrence rate of the dynamic cardiac ultrasound image, or to determine the gradient clipping threshold based on the analysis accuracy of the convolutional neural network model, and to determine the transmission power of the ultrasound probe based on the signal-to-noise ratio of the dynamic cardiac ultrasound image.

[0038] Specifically, dynamic cardiac ultrasound images include two-dimensional cardiac ultrasound images, three-dimensional cardiac ultrasound images, and Doppler blood flow images.

[0039] Specifically, preprocessing includes noise reduction, image enhancement, and image resizing.

[0040] Specifically, the optimized images include denoised two-dimensional cardiac ultrasound images, image-enhanced three-dimensional cardiac ultrasound images, and standardized Doppler blood flow images.

[0041] Specifically, stochastic gradient descent is an optimization algorithm used to minimize an objective function, thereby updating model parameters.

[0042] Specifically, dynamic features of cardiac ultrasound images include ventricular volume, ejection fraction, and valvular regurgitation velocity.

[0043] Specifically, artifacts are false images in an image that are unrelated to real anatomical structures or blood flow.

[0044] Specifically, the analysis results include cardiac structural assessment, functional parameter measurement, and hemodynamic analysis.

[0045] Specifically, the artifact rate of dynamic cardiac ultrasound images is the ratio of the number of artifacts to the total number of examinations.

[0046] Specifically, the pulse repetition frequency of an ultrasonic probe is the number of times the ultrasonic probe emits periodic pulse signals per unit time.

[0047] Specifically, the gradient clipping threshold is a limit value used to restrict the gradient size during the optimization of stochastic gradient descent.

[0048] Specifically, the signal-to-noise ratio of a dynamic cardiac ultrasound image is the ratio of the signal intensity of the real anatomical structure to the signal intensity of the noise.

[0049] Specifically, the transmission power of an ultrasonic probe is the energy output per unit time when the probe emits ultrasonic waves.

[0050] In implementation, the system of this invention, by setting up an image processing module, a model training module, a feature extraction module, and a control module, adjusts the pulse repetition rate according to the artifact occurrence rate of dynamic cardiac ultrasound images. Utilizing artificial intelligence image recognition training, it quickly identifies morphological changes in the heart under a four-chamber ultrasound view during cardiac arrest, as well as common cardiac lesions. Simultaneously, it intelligently identifies images with non-standard sections and prompts the scanner to adjust the scanning angle. For artifacts caused by air, bone, and other factors that affect the clarity and edge recognition of cardiac structures, increasing the duty cycle of the pulse repetition frequency increases the effective signal emission per unit time, thereby enhancing the intensity and continuity of the ultrasound echo signal. This reduces image fluctuations caused by signal attenuation, especially in areas with greater detection depth or weak tissue echoes. Stable signal output improves image contrast and resolution, reducing artifacts caused by signal instability. The gradient clipping threshold is adjusted based on the analysis accuracy of the convolutional neural network model. Due to insufficient generalization ability of the model, the convolutional neural network model may perform well on the training set. However, when faced with unseen real data, the accuracy and stability of feature extraction significantly decrease. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general patterns, leading to deviations in the extraction of the contours of fine cardiac structures. This, in turn, reduces the accuracy of the diagnostic assistance system. By reducing the gradient clipping threshold, excessively large gradients can be more strictly limited, suppressing the model's overfitting to noise and local features in the training data and preventing the model from misjudging occasional noise signals as key patterns. This reduces the deviation in the extraction of the contours of fine cardiac structures. Adjusting the transmission power of the ultrasound probe based on the signal-to-noise ratio of the dynamic cardiac ultrasound image is necessary. However, due to hardware limitations of the ultrasound equipment itself, image resolution is reduced, making fine cardiac structures such as chordae tendineae and valve edges blurry. Continuous and accurate spatiotemporal features cannot be obtained, making it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately resulting in a significant decrease in the accuracy of the diagnostic assistance system. Increasing the transmission power of the ultrasound probe can increase the initial energy of the ultrasound waves, enhancing the echo signal intensity of deep tissues and far-field structures, and improving image contrast.

[0051] Specifically, the control module is used to determine whether the processing stability of the dynamic cardiac ultrasound image meets the requirements based on the fact that the artifact occurrence rate of the dynamic cardiac ultrasound image is less than or equal to a preset first occurrence rate.

[0052] The control module is used to determine that the processing stability of the dynamic cardiac ultrasound image does not meet the requirements if the artifact occurrence rate of the dynamic cardiac ultrasound image is greater than a preset first occurrence rate.

[0053] Specifically, the control module is used to preliminarily determine that the training effectiveness of the convolutional neural network model is at risk of not meeting the requirements if the artifact occurrence rate of the cardiac ultrasound dynamic image is greater than the preset first occurrence rate and less than or equal to the preset second occurrence rate, and to determine whether the training effectiveness of the convolutional neural network model meets the requirements based on the analysis accuracy of the convolutional neural network model.

[0054] It is understandable that the three intervals defined by the preset first incidence rate and the preset second incidence rate correspond to three different scenarios:

[0055] The first interval is when the artifact incidence rate of the dynamic cardiac ultrasound image is less than or equal to the preset first incidence rate, which corresponds to the condition that the processing stability of the dynamic cardiac ultrasound image meets the requirements.

[0056] The second interval is when the artifact incidence rate of dynamic cardiac ultrasound images is greater than the preset first incidence rate and less than or equal to the preset second incidence rate. The corresponding situation is that due to the insufficient generalization ability of the model, the convolutional neural network model performs well on the training set, but when faced with unseen real data, the accuracy and stability of feature extraction decrease significantly. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general rules. The extraction of the contours of the heart's fine structures deviates, which in turn reduces the accuracy of the diagnostic assistance system.

[0057] The third interval is where the artifact incidence rate of dynamic cardiac ultrasound images is greater than the preset second incidence rate. The corresponding situation is: using artificial intelligence image recognition training, quickly identifying morphological changes of the heart in the four-chamber view of ultrasound during cardiac arrest, as well as common cardiac lesions. At the same time, intelligently identifying images with non-standard sections and prompting the scanner to adjust the scanning angle, and addressing artifacts caused by factors such as air and bones that affect the clarity of cardiac structure and edge recognition.

[0058] Understandably, the preset first incidence rate and preset second incidence rate can be set according to actual working conditions. These preset rates aim to ensure the accuracy and usability of the test results. Optionally, the preset first incidence rate and preset second incidence rate are adjusted based on the clinical impact of artifacts by extracting a large number of historical examination records from a clinical database, statistically analyzing the actual incidence rate of various artifacts, and using this as a benchmark. For example, the preset first incidence rate is typically selected within the range of [6%, 10%], and the preset second incidence rate is typically selected within the range of [11%, 15%].

[0059] Preferably, the first occurrence rate is 8% in the preferred embodiment, and the second occurrence rate is 13% in the preferred embodiment.

[0060] Specifically, the control module is used to increase the pulse repetition frequency of the ultrasound probe based on the fact that the artifact occurrence rate of the cardiac ultrasound dynamic image is greater than the preset second occurrence rate.

[0061] The increase in the pulse repetition frequency of the ultrasound probe is determined by the difference between the artifact occurrence rate of the dynamic cardiac ultrasound image and a preset second occurrence rate.

[0062] Specifically, when the difference between the artifact incidence rate of the dynamic cardiac ultrasound image and the preset second incidence rate is within 2%, the pulse repetition frequency of the ultrasound probe is increased to 1.1 times the original value. When the difference between the artifact incidence rate of the dynamic cardiac ultrasound image and the preset second incidence rate exceeds 2%, in addition to increasing to 1.1 times the original value, the pulse repetition frequency of the ultrasound probe is increased by 1 kHz for every 1% exceeding the preset second incidence rate. For example, if the difference between the artifact incidence rate of the dynamic cardiac ultrasound image and the preset second incidence rate is 4%, and the current pulse repetition frequency of the ultrasound probe is 10 kHz, the increased pulse repetition frequency of the ultrasound probe will be 10 × 1.1 + 2 × 1 = 13 kHz.

[0063] Specifically, kHz is the unit of pulse repetition frequency of an ultrasonic probe, which means kilohertz.

[0064] In practice, the system of this invention adjusts the pulse repetition rate by setting a preset first occurrence rate and a preset second occurrence rate. Utilizing artificial intelligence image recognition training, it quickly identifies morphological changes in the heart under a four-chamber ultrasound view during cardiac arrest, as well as common cardiac lesions. Simultaneously, it intelligently identifies images with non-standard sections and prompts the scanner to adjust the scanning angle. For artifacts caused by air, bone, and other factors that affect the clarity and edge recognition of the cardiac structure, increasing the duty cycle of the pulse repetition frequency increases the effective signal emission per unit time, thereby enhancing the intensity and continuity of the ultrasound echo signal. This reduces image fluctuations caused by signal attenuation, especially in areas with greater detection depth or weaker tissue echoes. Stable signal output improves image contrast and resolution, reduces artifacts caused by signal instability, and further enhances the processing stability of dynamic cardiac ultrasound images.

[0065] Specifically, the control module is used to determine that the training effectiveness of the convolutional neural network model meets the requirements based on the fact that the analysis accuracy of the convolutional neural network model is greater than or equal to a preset second accuracy.

[0066] The control module is used to determine that the training effectiveness of the convolutional neural network model does not meet the requirements if the analysis accuracy of the convolutional neural network model is less than a preset second accuracy.

[0067] Specifically, the control module is used to reduce the gradient clipping threshold based on the fact that the analysis accuracy of the convolutional neural network model is greater than a preset first accuracy and less than a preset second accuracy.

[0068] Specifically, the control module is used to preliminarily determine that the accuracy of extracting features from the dynamic cardiac ultrasound image is at risk of not meeting the requirements if the analysis accuracy of the convolutional neural network model is less than or equal to the preset first accuracy, and to determine whether the accuracy of extracting features from the dynamic cardiac ultrasound image meets the requirements based on the signal-to-noise ratio of the dynamic cardiac ultrasound image.

[0069] It is understandable that the three intervals defined by the preset first accuracy rate and the preset second accuracy rate correspond to three different scenarios:

[0070] The first interval is when the analysis accuracy of the convolutional neural network model is greater than or equal to the preset second accuracy, which corresponds to the situation where the training effectiveness of the convolutional neural network model is determined to meet the requirements.

[0071] The second interval is when the analysis accuracy of the convolutional neural network model is greater than the first preset accuracy but less than the second preset accuracy. The corresponding situation is that due to the insufficient generalization ability of the model, the convolutional neural network model performs well on the training set, but when faced with unseen real data, the accuracy and stability of feature extraction decrease significantly. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general rules. This leads to deviations in the extraction of the contours of the heart's fine structures, which in turn reduces the accuracy of the diagnostic assistance system.

[0072] The third interval is when the analysis accuracy of the convolutional neural network model is less than or equal to the preset first accuracy. The corresponding situation is that due to hardware problems of the ultrasound equipment itself, the image resolution is reduced, and the fine structures of the heart, such as chordae tendineae and valve edges, become blurred. It is impossible to obtain continuous and accurate spatiotemporal features, making it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately leading to a significant decrease in the accuracy of the diagnostic assistance system.

[0073] Understandably, the preset first accuracy rate and preset second accuracy rate can be set according to actual working conditions. These preset accuracy rates aim to ensure the accuracy and usability of the test results. Optionally, the preset first accuracy rate and preset second accuracy rate are determined by first establishing a target accuracy rate range based on task requirements and clinical standards, then obtaining reference benchmarks through literature review and baseline model testing, further optimizing the preset values ​​through cross-validation and confusion matrix analysis, and finally iteratively validating them in real clinical data. For example, the preset first accuracy rate is typically selected within the range of [89%, 93%], and the preset second accuracy rate is typically selected within the range of [94%, 98%].

[0074] Preferably, the first accuracy rate is 91% in the preferred embodiment, and the second accuracy rate is 95% in the preferred embodiment.

[0075] Specifically, the reduction in the gradient clipping threshold is determined by the difference between the analysis accuracy of the convolutional neural network model and a preset first accuracy.

[0076] Specifically, when the difference between the analysis accuracy of the convolutional neural network model and the preset first accuracy is within 2%, the gradient clipping threshold is reduced to 0.8 times the original value. When the difference between the analysis accuracy of the convolutional neural network model and the preset first accuracy exceeds 2%, the gradient clipping threshold is reduced by 0.5 for every 1% exceeding the original value, in addition to being reduced to 0.8 times the original value. For example, if the difference between the analysis accuracy of the convolutional neural network model and the preset first accuracy is 3%, and the current gradient clipping threshold is 5.0, the reduced gradient clipping threshold will be 5.0 × 0.8 - 0.5 × 1 = 3.5.

[0077] In implementation, the system of the present invention adjusts the gradient clipping threshold by setting a preset first accuracy rate and a preset second accuracy rate. Due to insufficient generalization ability of the model, the convolutional neural network model performs well on the training set, but when faced with unseen real data, the accuracy and stability of feature extraction decrease significantly. The model cannot effectively identify key patterns in the data and is prone to misjudging noise or local features in the training data as general rules, resulting in deviations in the extraction of the contours of the heart's fine structures. This leads to a decrease in the accuracy of the diagnostic assistance system. By reducing the gradient clipping threshold, excessively large gradients can be more strictly limited, suppressing the model's overfitting of noise and local features in the training data, and preventing the model from misjudging occasional noise signals as key patterns. This reduces the deviation in the extraction of the contours of the heart's fine structures and further improves the processing stability of dynamic cardiac ultrasound images.

[0078] Specifically, the control module is used to determine that the accuracy of extracting features from the dynamic cardiac ultrasound image does not meet the requirements if the signal-to-noise ratio of the dynamic cardiac ultrasound image is less than a preset signal-to-noise ratio, and to increase the transmission power of the ultrasound probe.

[0079] It is understandable that the two intervals of the preset signal-to-noise ratio correspond to two different situations:

[0080] The first interval is when the signal-to-noise ratio of the dynamic cardiac ultrasound image is greater than or equal to the preset signal-to-noise ratio, which corresponds to the situation where the accuracy of feature extraction of the dynamic cardiac ultrasound image meets the requirements.

[0081] The second interval is when the signal-to-noise ratio of the dynamic cardiac ultrasound image is greater than the preset signal-to-noise ratio. The corresponding situation is that due to hardware problems of the ultrasound equipment itself, the image resolution is reduced, and fine cardiac structures such as chordae tendineae and valve edges become blurred. It is impossible to obtain continuous and accurate spatiotemporal features, making it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately leading to a significant decrease in the accuracy of the diagnostic assistance system.

[0082] Understandably, the preset signal-to-noise ratio (SNR) can be set according to actual working conditions, aiming to ensure the accuracy and practicality of the test results. Optionally, the preset SNR is obtained by experimentally measuring the baseline SNR of typical clinical images, calculating the signal mean and noise standard deviation in the myocardial region and anechoic region respectively, to obtain the original SNR distribution. Subsequently, in conjunction with clinical experts, an initial threshold is determined according to diagnostic needs. Then, through simulation testing, controllable Gaussian noise is added to the original image to verify the detection rate of key lesions under the preset SNR. Finally, the final SNR is determined by combining equipment performance and patient-specific factors. For example, the preset SNR is generally selected in the range of [20dB, 25dB].

[0083] Preferably, the preset signal-to-noise ratio is 23dB in the preferred embodiment.

[0084] Specifically, the increase in the transmission power of the ultrasound probe is determined by the difference between the preset signal-to-noise ratio and the signal-to-noise ratio of the dynamic cardiac ultrasound image.

[0085] Specifically, when the difference between the signal-to-noise ratio (SNR) of the dynamic cardiac ultrasound image and the preset SNR is within 2 dB, the transmission power of the ultrasound probe is increased to 1.2 times the original value. When the difference between the SNR of the dynamic cardiac ultrasound image and the preset SNR exceeds 2 dB, in addition to increasing to 1.2 times the original value, the transmission power of the ultrasound probe is increased by 2 mW for every 1 dB exceeding the preset SNR. For example, when the difference between the SNR of the dynamic cardiac ultrasound image and the preset SNR is 3 dB, the current transmission power of the ultrasound probe is 30 mW, and the increased transmission power of the ultrasound probe is 30 × 1.2 + 2 × 1 = 38 mW.

[0086] Specifically, mW is the unit of transmission power of an ultrasonic probe, which means milliwatt.

[0087] In practice, the system described in this invention adjusts the transmission power of the ultrasound probe by setting a preset signal-to-noise ratio. Due to hardware limitations of the ultrasound equipment itself, the image resolution is reduced, and fine cardiac structures such as chordae tendineae and valve edges become blurred, making it impossible to obtain continuous and accurate spatiotemporal features. This makes it difficult for artificial intelligence to accurately extract their contours and morphological features, ultimately leading to a significant decrease in the accuracy of the diagnostic assistance system. By increasing the transmission power of the ultrasound probe, the initial energy of the ultrasound waves can be increased, thereby enhancing the echo signal intensity of deep tissues and far-field structures, improving image contrast, and further improving the processing stability of dynamic cardiac ultrasound images.

[0088] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

Claims

1. An artificial intelligence-based cardiac ultrasound examination assistance system, characterized by, The method comprises the following steps: An image processing module comprises an acquisition unit configured to acquire a cardiac ultrasound dynamic image using an ultrasound probe and a preprocessing unit connected to the acquisition unit and configured to preprocess the cardiac ultrasound dynamic image to output an optimized image; A model training module connected to the image processing module comprises a model generation unit configured to train an initial model according to the optimized image to output a convolutional neural network model and a model update unit connected to the model generation unit and configured to update the convolutional neural network model by a stochastic gradient descent optimization algorithm; A feature extraction module connected to the model training module comprises a feature extraction unit configured to extract key features of a cardiac structure by the convolutional neural network model to output cardiac ultrasound dynamic image features and an analysis unit connected to the feature extraction unit and configured to analyze the cardiac ultrasound dynamic image features to output an analysis result; A control module connected to the image processing module, the model training module, and the feature extraction module respectively is configured to determine a pulse repetition frequency of the ultrasound probe according to an artifact occurrence rate of the cardiac ultrasound dynamic image, and configured to determine that a processing stability of the cardiac ultrasound dynamic image meets a requirement according to the artifact occurrence rate of the cardiac ultrasound dynamic image being less than or equal to a preset first occurrence rate; The control module is configured to determine that the processing stability of the cardiac ultrasound dynamic image does not meet the requirement according to the artifact occurrence rate of the cardiac ultrasound dynamic image being greater than the preset first occurrence rate, or to determine a gradient clipping threshold according to an analysis accuracy of the convolutional neural network model; The control module is configured to preliminarily determine that a training effectiveness of the convolutional neural network model does not meet the requirement according to the artifact occurrence rate of the cardiac ultrasound dynamic image being greater than the preset first occurrence rate and less than or equal to a preset second occurrence rate.

2. The artificial intelligence-based cardiac ultrasound examination assistance system according to claim 1, characterized in that, The control module is configured to increase the pulse repetition frequency of the ultrasound probe according to the artifact occurrence rate of the cardiac ultrasound dynamic image being greater than the preset second occurrence rate; An increase amplitude of the pulse repetition frequency of the ultrasound probe is determined by a difference between the artifact occurrence rate of the cardiac ultrasound dynamic image and the preset second occurrence rate.

3. The artificial intelligence-based cardiac ultrasound examination assistance system according to claim 2, characterized in that, The control module is configured to determine that the training effectiveness of the convolutional neural network model meets the requirement according to the analysis accuracy of the convolutional neural network model being greater than or equal to a preset second accuracy; The control module is configured to determine that the training effectiveness of the convolutional neural network model does not meet the requirement according to the analysis accuracy of the convolutional neural network model being less than the preset second accuracy.

4. The artificial intelligence-based cardiac ultrasound examination assistance system according to claim 3, characterized in that, The control module is configured to decrease the gradient clipping threshold according to the analysis accuracy of the convolutional neural network model being greater than a preset first accuracy and less than the preset second accuracy.

5. The artificial intelligence-based cardiac ultrasound examination assistance system according to claim 4, characterized in that, The control module is configured to preliminarily determine that an extraction accuracy of the cardiac ultrasound dynamic image features does not meet the requirement according to the analysis accuracy of the convolutional neural network model being less than or equal to the preset first accuracy.

6. The artificial intelligence-based cardiac ultrasound examination assistance system according to claim 5, characterized in that, A decrease amplitude of the gradient clipping threshold is determined by a difference between the analysis accuracy of the convolutional neural network model and the preset first accuracy.

Images